Model

A model is the component of an AI system that has been trained to take inputs and produce outputs — a fraud score, a customer churn prediction, a document summary, a generated response. It is built through a training process that exposes it to large amounts of data, allowing it to learn patterns rather than follow explicitly coded rules. The model itself is distinct from the system around it: the data pipelines feeding it, the application presenting its outputs, the human review process downstream. When an AI system fails, the problem is often not the model but something in that surrounding infrastructure.

Accountability for an AI system's outputs requires knowing what the model is, who controls it, what it was trained on, and where its performance has limits. Organizations that treat the model as a black box — purchasing capabilities without understanding what they're running — lose the ability to explain decisions, diagnose failures, or respond to regulatory questions. The model is also where vendor dependency concentrates: if a provider changes, deprecates, or degrades a model, the organization needs to know what they're relying on and what it would take to replace it.